Home

Search

Collections

Journals

About

Contact us

My IOPscience

Broyden's method for self-consistent field convergence acceleration

This content has been downloaded from IOPscience. Please scroll down to see the full text.
1984 J. Phys. A: Math. Gen. 17 L317
(http://iopscience.iop.org/0305-4470/17/6/002)
View the table of contents for this issue, or go to the journal homepage for more

Download details:
IP Address: 128.118.88.48
This content was downloaded on 08/06/2014 at 20:50

Please note that terms and conditions apply.

J. Phys. A: Math. Gen. 17 (1984) L317-L321. Printed in Great Britain

LETTER TO THE EDITOR

Broyden's method for self -consistent field convergence

acceleration
G P Srivastava
Physics Department, New University of Ulster, Coleraine, N Ireland BT52 lSA, UK
Received 7 February 1984
Abstract. Broyden's method for self-consistent field convergence acceleration, recently
used by Bendt and Zunger in the band-structure calculation of solids, is reviewed and a
simpler computational scheme is presented.

Recently, Bendt and Zunger (1982) have derived a variational principle for obtaining
the minimum of the density-functional total energy of solids and molecules. For their
variational principle to be equivalent to the variational principle derived by Kohn and
Sham (1965) these authors have obtained a sufficient condition:
F ( p )= v - p = 0,

(1)

where p is an independent vector variable representing the electronic screening

potential and v = v(p) is the output screening potential to be determined from p.
Equation (1) is also the condition for the self-consistent band-structure calculation of
the system for which the total energy is required.
The system of simultaneous nonlinear equations F = O in equation (1) provides a
cherished topic in numerical analysis (see e.g. Byrne and Hall 1973). Of several
iterative methods available to solve systems of nonlinear equations, the most sophisticated one is Broyden's updated version of the quasi-Newton-Raphson iterative procedure (see, for example, Ralston and Rabinowitz 1978). In this procedure one
approximates F(p ) at various iterations, writing for the mth iteration
F ( p )=F(p"') +J"'(p -p"').
(2)
Further, one assumes that for the left-hand side to vanish the right-hand side must
also vanish, which yields the following quasi-Newton-Raphson method for the desired
solution
p(m+l) =p(m)-

( m ) -1

1 Fb"').

(3)
It is very well known that the success of this iterative procedure depends entirely on
how close the initial approximation for the Jacobian J'" = aF/ap"' is to the true
solution: if the initial approximation is not quite close to the solution, often the iteration
will not converge. Finding a good initial approximation to the Jacobian is a serious
problem. For p = { p l,p 2 ,. . . , p N } ,with a large value of N (which is usually the case
in the context of the present discussion), an explicit evaluation and inversion of the
initial Jacobian may itself involve excessive computational effort and storage requirements.
[J

0305-4470/84/060317+05$02.25 @ 1984 The Institute of Physics

L317

L318

Letter to the Editor

Broyden's modification (Broyden 1965) of the quasi-Newton-Raphson iterative

procedure is extremely helpful in achieving a fast convergence rate with a moderately
good initial guess for the Jacobian. According to this scheme one generates an
approximation $ ( I ) to the initial Jacobian J ( l ) and for m > 1 it is required that 9 ( m )
satisfies the following two conditions

(9

g = ~ ( m ) - ~ ( m - l ) $- ( m ) ( P ( m ) - P ( m - l ) )

(ii)

Q = I19(m)-9(m-1)112
be minimised.

=o

(4)

(5)

The consideration (ii), subject to (i), can be made by using the method of Lagrange
multipliers. Consider a functional
K =Q+ATg,

(6)

where A is a Lagrange multiplier vector. Then the variational principle aK/a9(" = 0

gives
$(m) =$(m-l)+tAT(P(m)-P(m-l)
).
(7)
When compared with equation (4), this gives the desired result for A, which when
used in equation (7) yields the following updating procedure for $ ( m )

This is called a rank-one updating procedure, as the rank of $ ( m ) - 9(m-1)

is one. Thus
equation (3) is used with equation (8) to achieve a fast rate of iterative convergence.
The initial guess 9 ' l ) can be made in accordance with the problem at hand. In the
present context of band-structure calculation, it can be shown that the Jacobian is the
negative of the dielectric kernel. Thus a simple guess for 9")may be the diagonal
form of the Thomas-Fermi screening
9 ( " =-( 1 + k 2 / G 2 ) f =
-( l / a (G ) ) f

(9)

k, being the Fermi momentum, and G is a reciprocal lattice

where k =
vector. An even simpler guess would be a diagonal constant matrix
= -( 1/
a)i
(10)
where 0 < a < 1. The particular choices in equations (9) and (10) without the updating
scheme of equation (8) are, respectively, Kerker's scheme (Kerker 1981) and the
potential mixing scheme (Bendt and Zunger 1982, Dunlap 1982, Dederichs and Zeller
1983). The choice of a in equations (9) and (10) should be such that Z E 1 [Fj2)I2<
I;, [Fj"I2 is satisfied (Powell 1970).
Bendt and Zunger have used Broyden's scheme to achieve rapid self-consistency
in the band-structure calculation of a few semiconductors. However, in many applications, when N is very large, the updating scheme of equation (8) is not desirable on
two counts: (i) storage of N 2 elements of 9 ( mmay
) be excessive, and furthermore,
.
both
(ii) the use of equation (3) requires inverting the huge matrix 9 ( m )Fortunately,
these problems can be dealt with if the following computational scheme is adopted.
).
using Broyden's second method, the updating
Let F?g(m)be the inverse of 9 ( m Then
formula for F ? ( ~ )is

Letter to the Editor

This equation, with the help of equation (9) or ( l o ) , can be expressed as
@ I ) = D = -pf

With this equation (3) can be expressed as the following vector equation
p(m+l) =p(m)+l)(m),

where
l)(m)

=p F ( m ) -

cmiu(i).

i=2

In equations (13) and ( 1 5) we have

Thus the iterative solution of F ( p ) = O can be obtained by using Broyden's method

which in the present scheme involves only storing a few vectors of length N.
To test the convergence rate of Broyden's method we have calculated the electronic
band structure and the total crystal energy of InP using a local version of the ab initio
pseudo-potentials of Zunger and Cohen (1978). Exchange and correlation were treated
in the local density formalism using the formulation of Hedin and Lundqvist (1971).
Figures 1( a ) , ( b )show the results of convergence test for, respectively, the screening
potential V,, and the total energy E,,, with the fixed Jacobian schemes of equations
(9) and (10). It is clearly seen that both V,, and E,,, diverge if no mixing of input
and output potentials is used: i.e. when 9 ( m ) = 9") = -f The results start to converge,
though slowly, even for a small mixing of input and output potentials. In the present
case of study it is found that a =0.5 (a 50%-50% mixing of the potentials) gives a
very fast convergence. The Thomas-Fermi approximation for the Jacobian gives a
better convergence rate than using the simple mixing case with a = 0.8 (20% input
mixing) but shows somewhat slower convergence rate than using a = 0.5. Very recently,
Dederichs and Zeller ( 1983) have rigorously analysed the simple mixing procedure
and have obtained limits on a for this procedure to converge. An optimum value of
a may depend on how good the starting input potential p is.
Broyden's method of updating the Jacobian was found to accelerate the convergence
rates in both V,, and E,,, for all the cases studied in figure 1 (except a = 1). As can
be inferred from equations (14) and (15), acceleration of the convergence will start
at the third iteration. In figure 2 we show the comparison of Broyden's method with
the potential mixing schemes using the fixed Jacobians 9g)and 9 ( m ) ( a = 0.8). For
a stability of
Ryd in E,,, Broyden's method saves three iterations in this 'easy'
case of two atoms/cell. For difficult cases (e.g. in slab surface calculations) this method
must prove very time saving on the computer. However, if a value of a has been

L320

Letter to the Editor

-1720'

Fig. 1. Convergence studies on ( a ) the screening

potential and (b) the total energy of InP using the
fixed Jacobian scheme. Curves 1-4use equation (10)
with a = 1.0, 0.95,0.8 and 0.5 respectively. Curve
5 uses equation (9)

\
\

lb)

1
F i r e 2. Convergence acceleration using Broyden's
method. Curves 2 and 4 use equation (10) with
a =0.8 and equation (9), respectively, for 4"'.
Curves 1 and 3 correspond to curves 2 and 4 without
Jacobian update.

found for which the simple mixing scheme converges rapidly, then Broyden's method
may not offer much saving in computer time. Such is the case with 9")( a= 0.5) in
the present study, where Broyden's method saves only one (two) iteration(s) for a
convergence of
Ryd in E,,,.
We believe that Broyden's method is superior to the recently proposed perturbative
approach of Dunlap and the accelerated schemes presented in Dederichs and Zeller.
Pulay (1980, 1982) has proposed an alternative approach for accelerating the convergence rate of the quasi-Newton-Raphson procedure.
It is a great pleasure to acknowledge very helpful discussions with Dr A Zunger and
Professor P Pulay.

References
Bendt P and Zunger A 1982 Phys. Rev. B 26 3114
Broyden C G 1965 Marh. Compur. 19 577
Byrne G and Hall C A 1973 Numerical Solurion of Sysrems of Nonlinear Algebraic Equations (New York:
Academic)
Dederichs P H and Zeller R 1983 Phys. Rev. B 28 5462

Letter to the Editor

L321

Dunlap B I 1982 Phys. Rev. A 25 2847

Hedin L and Lundqvist B I 1971 J. Phys. C: Solid State Phys 4 2064
Kerker G P 1981 Phys. Rev. B 23 3082
Powell M J D 1970 Numerical Methods for Nonlinear Algebraic Equations ed P Rabinowitz (New York:
Gordon and Breach) p 87
Pulay P 1980 Chem. Phys. Len. 73 393
1982 J. Compt. Chem. 3 556
Ralston A and Rabinowitz P 1978 A First Course in Numerical Analysis, 2nd edn ch 8 (London: McGraw-Hill
Kogakusha)
Zunger A and Cohen M L 1978 Phys. Rev. B 18 5449